Creating a Model

In addition to using a public model, you can also create your own model when deploying an inference service on DataArtsFabric. You can create models on the Model page of DataArtsFabric. These models belong to you and are invisible to other users.

Notes and Constraints

The common constraints for creating a model are as follows:

The model you create must be a base model supported by DataArtsFabric. The following table lists the base models.

**Table 1** Base models
Base Model Type	Description
QWEN_2_72B	With 72 billion parameters, Qwen2 outperforms most previous open-weight models in multiple benchmark tests in terms of language understanding, generation, multilingual capabilities, coding, mathematics, and inference. It is also competitive with proprietary models.
GLM_4_9B	GLM-4-9B is an open-source version of the latest-generation pre-trained GLM-4 series models launched by Zhipu AI. With 9 billion parameters, it has high performance in datasets evaluation in terms of semantics, mathematics, inference, code, and knowledge.
LLAMA_3_8B	With 8 billion parameters, it is the third-generation model of the Llama series. Llama3 has achieved leading performance in multiple benchmark tests. The model uses a large amount of Chinese data for pre-training, expanding the coverage of the Chinese character sets.
LLAMA_3_70B	With 70 billion parameters, it is the third-generation model of the Llama series. Llama3 has achieved leading performance in multiple benchmark tests.
LLAMA_3.1_8B	Llama3.1 is the first publicly available model and performs well in common sense, steerability, mathematics, tool usage, and multilingual translation. It supports advanced use cases, such as long text summarization, multilingual dialogue agents, and coding assistants. With 8 billion parameters, the model uses a large amount of Chinese data for pre-training, expanding the coverage of the Chinese character sets.
LLAMA_3.1_70B	Llama3.1 is the first publicly available model and is close to top AI models in terms of common sense, steerability, mathematics, tool usage, and multilingual translation. It supports advanced use cases, such as long text summarization, multilingual dialogue agents, and coding assistants. The model has 70 billion parameters.

The model must be in safetensors format. The safetensors format developed by Huggingface is a reliable and easy-to-port storage format of machine learning models. It is used to securely store tensors at a high speed.

Click the following URLs to view the format requirements for specific model examples.

Base Model Type	Model Example Name	Model Source
LLAMA_3_8B	Llama 3 8B Chinese Instruct	https://www.modelscope.cn/models/FlagAlpha/Llama3-Chinese-8B-Instruct
LLAMA_3_70B	Llama 3 70B	https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct
LLAMA_3.1_8B	Llama 3.1 8B Chinese Chat	https://modelscope.cn/models/XD_AI/Llama3.1-8B-Chinese-Chat
LLAMA_3.1_70B	Llama 3.1 70B	https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct
QWEN_2_72B	Qwen 2 72B Instruct	https://huggingface.co/Qwen/Qwen2-72B
GLM_4_9B	Glm 4 9B Chat	https://huggingface.co/THUDM/glm-4-9b-chat

Prerequisites

You have a valid Huawei Cloud account.
You have at least one workspace available.

You have created an OBS bucket and folders for storing models, uploaded model files that meet the requirements, and ensured that the OBS bucket for storing models is in the same region as DataArtsFabric. For details, see Creating an OBS Bucket.

Procedure

Log in to Workspace Management Console.
Select the created workspace and click Access Workspace.
In the navigation pane on the left, choose Resources and Assets > Model. The Model page is displayed.
Click Create Model. The Create Model page is displayed.

Enter basic model information, including the name and description, select the OBS path of the model file, and click Create Now. For details, see the following table.

**Table 2** Basic information about the model to be created
Parameter	Description
Model Name	Indicates the model name, which is mandatory. The name contains 1 to 64 characters and must be unique. Only letters, digits, underscores (_), hyphens (-), periods (.), and spaces are allowed.
Model Description	Indicates the model description, which is optional. The value contains 0 to 1,024 characters. Special characters such as ^!<>=&"' are not supported.
Version Name	Indicates the version name, which is mandatory. The name contains 1 to 64 characters and must be unique. Only letters, digits, underscores (_), hyphens (-), periods (.), and spaces are allowed.
Version Description	Indicates the version description, which is optional. The value contains 0 to 1,024 characters. Special characters such as ^!<>=&"' are not supported.
Base Model Type	Indicates the base model type, which is mandatory. For details, see Table 1.
Model File Path	Indicates the model file path, which is mandatory. Currently, the OBS path is supported. The current user must have the read permission on the path.